Periodic Task Allocation Method Adapting to Unstable Resource Status on the MPI Enabled Distributed System

摘要

High performance distributed system is generally consists of large amount of computing resources to perform several tasks simultaneously. If performance of computing nodes in the distributed system is symmetric, load balancing is simply settled by assigning same amount of tasks to the computing nodes. On the other hand, if performance of computing nodes in the distributed system is asymmetric and unstable, load balancing has to be settled by considering such a system status. MPICH, which is used for communication among the nodes in the parallel system, makes it more difficult to assign tasks dynamically on the variable system environment since it requires all of computing nodes in the system to be synchronized to communicate with each other. In this paper, we introduce a mechanism to perform a job on the MPI enabled cluster system organized with workstations. It is proposed new load-balancing scheme considering the unstable computing and network resources on the MPI enabled cluster system. Experimental results show that we can reduce job completion time up to 30% in comparing to the previous method.